基于迁移学习的造纸断纸故障诊断方法研究

Methodology Research on Paper Breaking Fault Diagnosis Based on Transfer Learning

关键词:  迁移学习  断纸  故障诊断  方法研究  工况

Key Words:transfer learning  paper breaking  fault diagnosis  methodology research  working condition

摘要:本课题面向断纸故障标记数据短缺而生产工况切换较多，故障诊断建模复用困难的问题，提出了分别基于参数和特征的断纸故障迁移模型建模方法，通过分析定量设定值及其强相关变量的数据分布特征，对工业数据进行了基础工况划分，并以马氏距离、多核最大均值差异等距离函数评估验证了工况划分的可靠性。在所划分的工况数据基础上，将依据有效断纸故障数据较多的工况建立的断纸故障模型迁移至数据缺失的工况中。结果表明，建立的故障诊断迁移模型在不同的工况迁移任务中分别达到了98.3%、94.6%、96.4%的诊断准确率，提高了模型的泛用性，促进其面向不同的造纸生产过程进行更广泛和更精确的故障诊断。

Abstract:Aiming at the shortage of paper breaking fault marker data and the difficulty of reusing fault diagnosis modeling due to the frequent switching of production conditions， this paper proposed a modeling method of paper breaking fault migration model based on parameters and features， respectively. By analyzing the data distribution characteristics of quantitative setpoints and their strongly correlated variables， the basic working conditions of industrial data were divided. The reliability of the working condition division was verified by the evaluation of Mahalanobis distance and multi-core maximum mean difference equidistance function. Based on the divided working condition data， the paper breaking fault model established according to the working condition with more effective paper breaking fault data was transferred to the working condition with missing data. The results showed that the established fault diagnosis transfer model could achieve 98.3%， 94.6%， and 96.4% diagnostic accuracy in different working conditions， respectively， which improved the universality of the model and promoted the wider and more accurate fault diagnosis for different papermaking processes.